Research of the Correlation between the H1N1 Morbidity Data and Google Trends in Egypt

Shengli Li, Xichuan Zhou

Abstract

The search engine based on influenza monitoring system has been widely applied in many European and American countries. However, there are not any correlative researches reported for African developing countries. Especially, the countries Egypt has not designed an influenza monitoring system on the basis of the Internet search data. This study aims at analyzing the correlation between the Google search data and the H1N1 morbidity data of Egypt, and examining the feasibility of Google Flu Model in predicting the H1N1 influenza trend.

TL;DR

This paper investigates whether Google search data can serve as an early warning system for H1N1 outbreaks in Egypt, a region where traditional surveillance often lags by weeks. The researchers found that while search trends can correlate strongly with actual infection rates—achieving a correlation of 0.76 when search data leads official reports by two weeks—the model’s reliability fluctuates significantly based on media hype and changing public interest.

What Problem Does This Paper Address?

Traditional disease surveillance is slow. In many developing nations, including Egypt, official health reports on influenza (like H1N1) can take one to two weeks to compile. In the fast-moving context of a pandemic, even a seven-day delay can cripple prevention efforts.

While "Google Flu Trends" (GFT) proved successful in Western countries, its efficacy in African and Middle Eastern contexts remained largely untested. This paper addresses the "digital divide" in epidemiology, exploring whether the search behaviors of Egyptian internet users provide enough signal to predict morbidity trends in a country with different linguistic, economic, and cultural profiles than the US or Europe.

Methods

The authors employed a multi-step statistical approach to bridge the gap between search queries and clinical data:

  1. Query Selection: They identified 50 potential search terms (30 in English, 20 in Arabic) related to flu symptoms, treatments (like Tamiflu), and names (like "pig flu").
  2. The Greedy Method: To avoid "overfitting" the model with useless data, they used a Greedy Method to select the optimal combination of queries that yielded the highest correlation with actual H1N1 cases.
  3. Time-Shifting Analysis: The researchers tested the data for "lagging" and "preceding" effects. In their terminology, "lagging" meant shifting morbidity data behind search data to see if today’s searches could predict cases two weeks from now.
  4. Linear Regression: They utilized a standard least-squares regression model to estimate the relationship between the normalized search volume (0-100 scale) and the number of H1N1 cases reported by the WHO.

Datasets

The study analyzed a five-year window from January 4, 2009, to December 31, 2013 (261 weeks).

  • Morbidity Data: Weekly H1N1 case numbers for Egypt sourced from the World Health Organization (WHO) website, specifically provided by the Global Influenza Surveillance Response System (GISRS).
  • Search Data: Weekly search volume data for the selected 50 queries, filtered by the location parameter "Egypt," sourced directly from Google Trends.

Key Insights and Ideas

  • The "Media Effect" vs. Reality: One of the most striking findings was the behavior of the query "Tamiflu." It peaked sharply in April 2009 following news reports of the outbreak in Mexico, even though there were very few actual cases in Egypt at that time. This suggests that search data often reflects anxiety or curiosity rather than actual infection.
  • Language is Local: Arabic queries were essential. Terms like "search swine flu" in Arabic showed a higher correlation (0.58) than many English counterparts, highlighting the need for localized digital epidemiology.
  • The "Aha" Moment on Lagging: The model performed best when it was used to "predict" the future. The correlation coefficient jumped from 0.71 (real-time) to 0.76 when the search data was compared to morbidity data from two weeks later. This confirms that people turn to Google the moment they feel ill—well before they enter the official medical record.
Correlation coefficients for different time lags, showing peak correlation at a 2-week lag.
  • Diminishing Returns: As the pandemic moved into 2011–2013, the correlation dropped significantly. The authors suggest that once a disease is no longer "new," people stop searching for it even if they are sick, because they already know the symptoms and treatments.

Key Results

  • Peak Correlation: The highest overall correlation achieved was 0.76 (p < 0.05) when the H1N1 data lagged the search data by two weeks.
  • Annual Variance: In 2010, specific queries like "pig flu" (Arabic) reached a staggering 0.88 correlation.
  • Predictive Accuracy: While the model successfully tracked the timing of the peaks in late 2009 and early 2010, it struggled with the magnitude. It produced a "false alarm" in March 2010, predicting a surge that never materialized in the official WHO data.
Comparison of actual H1N1 cases (black) vs. search-based model predictions (red).

Limitations and Future Work

The authors are transparent about the hurdles facing digital disease tracking in developing nations:

  • The Digital Divide: Unbalanced economic status and internet access in Egypt mean the data primarily reflects urban, wealthier populations.
  • Behavioral Shifts: Human behavior is a moving target. The "algorithms" of human curiosity change as people become educated about a virus, requiring the model to be constantly revised.
  • Causality vs. Correlation: The study identifies a mathematical link but does not prove that searching causes a prediction of illness (it could be driven by news cycles).
  • Future Directions: The authors suggest that future models must account for "noise" from media reports and should be expanded to other developing countries in Africa to create a more robust, real-time syndromic monitoring system.

Summary updated

Abstract

The search engine based on influenza monitoring system has been widely applied in many European and American countries. However, there are not any correlative researches reported for African developing countries. Especially, the countries Egypt has not designed an influenza monitoring system on the basis of the Internet search data. This study aims at analyzing the correlation between the Google search data and the H1N1 morbidity data of Egypt, and examining the feasibility of Google Flu Model in predicting the H1N1 influenza trend.
BibTeX
@misc{li2015research,
  author = {Shengli Li and Xichuan Zhou},
  title = {{Research of the Correlation between the H1N1 Morbidity Data and Google Trends in Egypt}},
  year = {2015},
  eprint = {1511.05300},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1511.05300},
}
Figure